AI Paper Atlas

Research map / Architectures and efficiency / CNN design

Visualizing Deep Convolutional Neural Networks Using Natural Pre-images

Aravindh Mahendran, Andrea Vedaldi

IJCV 2016 · Architectures and efficiency › CNN design · approach: Compact CNN architectures · 502 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
use the proposed regularized optimization instead of HOGgleHOG inversion · ILSVRC 2012 valreconstruction error (%): 60.1 -> 36.6Improved
use bilinear gradient orientation assignment as in SIFT (ablation)HOG inversion · ILSVRC 2012 valreconstruction error (%): 36.6 -> 11.5Improved
use stronger regularization (small C) (ablation)AlexNet inversion, early layers · ILSVRC 2012 valreconstruction qualityHurt
use stronger regularization (small C) (ablation)AlexNet inversion, deep layers · ILSVRC 2012 valreconstruction qualityNo clear effect
add at least weak regularization (C=100 or C=1) vs none (ablation)AlexNet pre-images · user study, 13 subjectsinterpretability preferenceImproved
use strong regularization (C=1) on deep layers (ablation)AlexNet pre-images · user study, 13 subjectsinterpretability preferenceImproved
replace TV (beta=1) with TV^beta (beta=1.5 or 2) (ablation)pre-image regularizer · AlexNet relu3 inversionimage qualityMixed
add jitter regularizer (ablation)activation maximization · AlexNet fc8 tree frog neuronvisual structureImproved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy